# Controlled GitHub admission proof for [#19795](https://github.com/pingdotgg/t3code/pull/19795)

Head: [540380bf8d](https://github.com/pingdotgg/t3code/commit/540380bf8d4c23bc09e375e1c7e10a58c84d75c5). Baseline: [1451e7cb23](https://github.com/pingdotgg/t3code/commit/1451e7cb23034d29774338608368a15d8845df36). Headless Linux, Node24. This report measures the production request-policy component over real loopback HTTP with a synthetic credential. It does not measure composer rendering, a running full T3 server, real GitHub latency, or native clients.

## Matched measurement

Identical dependencies and 40 ms response delay. Eight active held background requests, 24 queued background requests, then eight ordinary requests. Release the active responses after all callers are queued. Five alternating main/head/head/main cycles, ten samples per revision per condition, after idle warmups. Same40 requests per loaded sample, all once, at most eight active, complete drain. Median is the mean of the two middle samples. The main policy is the exact pinned Git blob with only two relative imports made absolute; hashes and all raw samples are in the JSON.

| Median | Main | Rebuilt head |
|---|---:|---:|
| Foreground completion under backlog | 241.25 ms | 137.93 ms |
| Background completion | 193.89 ms | 232.99 ms |
| All requests drained | 241.27 ms | 233.00 ms |
| Idle single foreground read | 42.14 ms | 42.24 ms |

The first ordinary HTTP request was index32 on main and9 on head in every loaded sample (zero-based). Background completion is slower because ordinary callers pass the queued backlog. These numbers describe one deliberately saturated fixture; they do not validate the historical22% composer claim, a universal fairness bound, or memory overhead.

## Generated summary API integration

A second fixture runs the production GitHubPullRequestApi, generated aliased resolver, GraphQL client, REST transport and shared request policy. Only external transport destination, credential and SourceControlHost are synthetic. The peer is a real loopback HTTP server. Eight held and eight queued background REST requests precede three ordinary summary reads. The actual GraphQL request starts at index9, ahead of the remaining background backlog. All17 HTTP requests drain, maximum active8. All preserve the fixture credential.

Three logical summaries stay one generated three-alias HTTP request. Healthy summaries7 and8 return their distinct branches. GitHub's alias-local NOT_FOUND for9 fails only9 with GitHubApiNotFoundError; healthy siblings are preserved. Fourteen assertions pass at this same head. The existing main error/caching suites and retained mixed host/credential/scope regression cases also pass.

## Focused checks

539 cases/eight original suites →498 cases/seven retained suites, comprising493 main baseline cases and five added cases. The original28 added declarations/46 expanded cases become three declarations/five cases in four existing suites plus one shared existing fixture. All1,446 baseline assertions remain. Exact-main policy/REST negative control fails the ordering regression (index16 versus <=10), followed by byte-verified restoration and14/14 REST cases passing.

All498 final-head cases pass, sequentially with nice10 and two Vitest workers. Scoped GitHub/core/server typechecks, targeted lint/format and diff check pass. Two pre-existing unused-parameter warnings remain. Credential settings snapshots, early quota-header observation, and duplicate error matrices were removed from this change.

## Limits and human verification

No fresh composer render/paint recording or visual claim. Earlier T3 preview attempts in this batch returned `Preview automation snapshot timed out after 15000ms.` and `Preview automation navigate timed out after 15000ms.` API and test-level evidence is used under the requested fallback. No real GitHub/GHE/GHES credential, rate-limit incident, macOS, Windows, iOS, Android or Electron desktop run was performed. The shared server path applies across clients and local/remote connections; that does not constitute native verification.

Before calling a composer percentage verified, use a real client with a fixed large linked-PR fixture, capture matched main/head foreground completion and background drain, and check warm/idle behavior. A foreground reader joining an already-running shared background cache producer is not promoted; the existing first-producer behavior is unchanged. The finite progress/cancellation tests do not prove all scheduler interleavings. Human approval of cross-cutting scheduling changes remains a repository review gate. Nothing was merged.

Model/harness: GPT-6 Astra tests, GPT-6.1 Sol implementation and integration proof through Codex in T3 Code; independent GPT-6 Sol and Claude Opus5.5 review reports are tracked separately.